Per Macrumors, on October 15 (local time), Meta introduced its cutting-edge open-weight model, Muse Spark Open, touted by the company as the most advanced open-weight model currently available in the United States. This model leverages a sophisticated Mixture of Experts (MoE) architecture, boasting a staggering 816 billion parameters in total, with 42 billion active parameters, and an impressive context window capable of handling up to 1 million tokens. It seamlessly integrates support for text, image, video, and audio inputs.

Optimized through the state-of-the-art MLX framework, Muse Spark Open is designed to operate locally on two Mac Studios, each outfitted with M5 Ultra chips and 512GB of unified memory. This setup is equivalent to harnessing the power of a supercomputer with 1TB of unified memory. Meta's internal testing reveals that the model excels in benchmarks such as JobBench and MCP-Atlas. However, it falls slightly short when compared to models like Kimi K3 and GLM 5.3 in areas such as terminal programming, software engineering tasks, and GDPVal-AA professional work capability evaluations.